AI for detecting and localizing concurrent abrupt and incipient leaks in water distribution networks

Water leakage detection and localization in water distribution networks (WDNs) are critical for sustainable water management, economic efficiency, and infrastructure integrity. Traditional inspection and hydraulic simulation methods often fail to provide timely and accurate results, prompting growing interest in hybrid approaches that combine model-based hydraulic simulations with data-driven machine learning techniques. This study proposes a novel hybrid method that leverages simulated SCADA data to train machine learning models for leakage detection and localization. Graph Transformers are trained to predict the pressure at each junction of the WDN based on observed pressure readings and auxiliary hydraulic features such as pump flow rates. The residual signal between predicted leakage-free and leaky pressures, along with its slope, is then used for detecting and localizing abrupt and incipient leakages, respectively. A residual signal compensation mechanism and a slope-based leakage detection and localization algorithm are introduced to address complex scenarios involving concurrent abrupt and incipient leakages - conditions frequently observed in real systems but rarely considered in existing learning-based hybrid methods. The proposed method was validated using benchmark datasets from the BattLeDIM2020 challenge and compared against state-of-the-art algorithms. Results show superior economic scores (€422,660±11,553), higher true positive rates, and improved robustness under multiple-leak conditions. These findings demonstrate the potential of advanced hybrid frameworks to significantly enhance the accuracy and applicability of leakage detection and localization in complex WDNs.